Use Plausibility Screens to Flag Extreme Records

Why Use Plausibility Screens?

A dose of 40,000 mg, a sample time of 240 hours, or a concentration far above the usual range can each deserve attention. But unusual is not the same as impossible, and an extreme value is not automatically a data error. A high concentration could be a real observation; a time far from the nominal visit could reflect the actual collection time. Conversely, a value that looks ordinary may still be wrong. A plausibility screen helps identify records that warrant review, without deciding their validity by itself.

The useful screen is specific to the field and its context. A dose rule should reflect the regimen and dose definition; a time rule should state which time reference it uses; a concentration rule should reflect the assay and units. Thresholds should be written down before applying them, with their rationale and action. This makes the review reproducible and reduces the risk of treating a statistical surprise as proof of an error.

Specify the Rule Before Applying It

A usable rule identifies the field, the context, the allowed values, the treatment of boundary values, and the action taken when a value falls outside the range. Context may include protocol, dose occasion, formulation, assay, study period, or whether a time is nominal or actual. The earlier material on nominal and actual sampling times, unit harmonization, and dose holds and restarts matters here: applying a threshold to the wrong time definition, unit, or dosing state can create misleading flags.

Write a rule in a form that could be applied by another analyst without guessing. For a measured field \(x\), a simple two-sided rule is:

Here \(c\) denotes the relevant context, and \(L(c)\) and \(U(c)\) are the lower and upper limits for that context. The rule makes the boundaries inclusive for values not flagged. A one-sided rule can be used when only one tail has a meaningful limit. If a field is missing, censored, or recorded as below quantification, handle it according to its own data definition; do not treat it as an ordinary numeric extreme.

The limits should come from defensible information: protocol constraints, source-system definitions, collection windows, known dose formats, or assay documentation. When no hard limit is supported, a screening range can still be chosen for review, but label it as a review threshold rather than a biological or clinical boundary. Record the rule's units, rationale, scope, and version, and state what happens to flagged records: typically they are queued for review and retained pending resolution.

Apply Screens to the Field They Describe

Dose records may represent an administered amount, a planned amount, or an amount per unit time. These are not interchangeable. Define the dose field first, then set a range appropriate to the regimen and dosing occasion. If protocol rules differ across groups or periods, use the appropriate context-specific range rather than forcing all records through one universal threshold.

Time screens likewise need a clear time origin and meaning. A negative elapsed time may indicate a problem with the time origin or record, but a sample collected later than its nominal visit is not necessarily invalid. Set a practical collection window from the study schedule and operational rules, and apply it to actual elapsed time when that is the field being screened. Preserve the nominal visit label separately; it is not a substitute for actual collection time.

For concentrations, distinguish a numeric reported result from missing or below-quantification records. A screen can flag values outside a documented assay reporting or validated range, or values beyond a deliberately broad review range. An assay's range may imply that the result needs investigation—for example, to check dilution or reporting metadata—but the screen alone does not determine whether the concentration is erroneous. Keep the concentration unit and assay context attached to the rule.

Why Unit and Boundary Rules Matter

A simple unit conversion preserves a screen's meaning only when the value and its limits are converted together. If \(y = ax\) for a positive conversion factor \(a\), then \(x > U\) is equivalent to \(ax > aU\). Thus a concentration of 26.4 mg/L exceeds 20 mg/L exactly when its converted value of 26,400 ng/mL exceeds 20,000 ng/mL. Comparing a converted value with an unconverted limit can produce a false flag or conceal a real one.

Boundary rules also prevent inconsistent decisions. For the inclusive range \(L \le x \le U\), values exactly equal to \(L\) or \(U\) pass the screen. If rounding is used for display, apply the rule to the stored value at its defined precision, not a rounded display value that can move a record across a threshold. State any precision or rounding convention explicitly.

Distribution-Based Screens: Useful but Limited

When a defensible hard range is unavailable, a distribution-based rule can help prioritize unusual records for review. One common exploratory rule uses quartiles: \(Q_1\) is the 25th percentile, \(Q_3\) is the 75th percentile, and the interquartile range is \(\mathrm{IQR} = Q_3 - Q_1\). The rule flags values below \(Q_1 - 1.5\mathrm{IQR}\) or above \(Q_3 + 1.5\mathrm{IQR}\). These are exploratory fences, not statistical tests of whether a value is erroneous.

The justification is a spread-based distance from the central half of the observed values. The interval from \(Q_1\) to \(Q_3\) has width \(\mathrm{IQR}\); extending each endpoint by \(1.5\) times that width creates outer review fences. Because quartiles are less sensitive than the mean and standard deviation to a few extreme values, this rule can be useful for triage. But the fences depend on the observed distribution and sample composition. Skewness, multiple subgroups, or a small sample can make them poor indicators of data quality.

Review, Diagnostics, and Failure Modes

A screen is working as intended when it identifies records for consistent, documented review. Before running it, check that the field is populated as expected, units have been harmonized, the correct contextual thresholds are attached, and missing or below-quantification records are not accidentally treated as numeric values. After running it, inspect the number and pattern of flags. A sudden surge can indicate a unit mismatch, an incorrectly applied threshold, or a change in source coding; a very small number of flags does not prove that the dataset is error-free.

A useful review record includes the raw value, the field and unit, the rule and version, the reason for the flag, the evidence considered, the disposition, and the person or process responsible for the decision. Preserve the raw value even if a later review supports a correction. Distinguish “flagged,” “reviewed,” and “corrected” as different statuses. Do not silently overwrite a value or remove a record merely because it exceeds a threshold.

The main alternatives differ in what they can establish. A protocol- or assay-based threshold is interpretable when its source and scope are explicit, but may not cover all contexts. A distribution-based rule can efficiently rank unusual observations, but can flag valid extremes or miss errors that are not statistically unusual. Manual review can consider context that a formula cannot, but needs consistent documentation to avoid arbitrary decisions. These approaches can complement one another; none turns a flag alone into proof of validity or error.

• Using one universal limit. Dose, time, and concentration fields have different meanings and units. Define separate rules and context-specific limits.
• Confusing a scheduled time with an actual time. A visit label does not replace the recorded collection time. State which field the rule evaluates.
• Using mismatched units. Convert both values and limits, and verify the conversion before interpreting a flag.
• Treating an outlier fence as a deletion rule. The fence marks unusualness relative to observed data; it does not prove an error.
• Removing flagged values automatically. A plausible screen identifies candidates for review. Retain the value unless a separate, documented decision supports correction or exclusion.